集成多场景的2D与3D融合人员检测方法研究OA
Multi-Scenario Occupant Detection and Recognition Based on 2D and 3D Data Fusion
针对智能座舱多场景检测需求,提出一种2D与3D数据融合检测方法,用于集成实现车内生命体遗留检测、分心驾驶和人脸身份识别功能.该方法将Stellar 200 3D相机采集的红绿蓝三通道图像(RGB)、红外(IR)和深度(Depth)数据作为输入,基于深度学习卷积神经网络模型开发出适用于不同场景的算法.选用MobileNet-V2作为分类网络,通过IR图像输入、IR-Depth融合输入训练模型,得到分心驾驶检测结果;通过YOLOv5s筛选人体目标后,经多任务卷积神经网络(MTCNN)进行人脸检测,再通过年龄分类卷积神经网络(Age-CNN)提取年龄特征,得到多个功能场景下的具体遗留结果;采用MTCNN算法实时检测视频流人脸,搭配轻量化人脸识别网络(MobileFaceNet)提取人脸高维特征嵌入,以实现精确的身份匹配.实车验证结果表明,该方案具有很高的准确性,且得益于多模态数据融合特性,能消除外界环境对检测效果的影响.
To meet the demand for multi-scenario detection in smart cockpits,this paper proposes a method that integrates 2D and 3D data fusion to achieve in-vehicle occupant detection,distracted driving recognition,and face recognition.The method uses three-channel Red-Green-Blue(RGB)images,Infrared(IR),and Depth data acquired by a Stellar 200 3D camera as input.Task-specific algorithms based on deep convolutional neural networks are developed for different scenarios.The MobileNet-V2 is employed as the classification network for distracted driving detection,and the model is trained using both IR images and fused IR-Depth inputs.For left-behind occupant detection,YOLOv5s first identifies human targets,Multi-Task Convolutional Neural Network(MTCNN)then detects faces,and Age Classification CNN(Age-CNN)extracts age features to determine the type of occupant left behind.For face recognition,the MTCNN algorithm detects faces in the video stream in real time,and the lightweight MobileFaceNet extracts high-dimensional facial embeddings for accurate identity matching.Real-vehicle validation results demonstrate the high accuracy of the proposed method.By fusing multimodal data,the method effectively reduces the impact of changes in the external environment on detection performance.
张海洋;杨云霄;曹立波;赵诚;安先龙;高超
湖南大学,长沙 410082湖南大学,长沙 410082湖南大学,长沙 410082中机科(北京)车辆检测工程研究院有限公司,北京 102100中机科(北京)车辆检测工程研究院有限公司,北京 102100中机科(北京)车辆检测工程研究院有限公司,北京 102100
信息技术与安全科学
车内生命体遗留分心驾驶人脸身份识别卷积神经网络
residual life forms in the cardistracted drivingfacial identificationconvolutional neural network
《汽车工程学报》 2026 (5)
770-778,9
湖南大学整车先进设计制造技术全国重点实验室自主研究课题资助项目(72475001)
评论